OpenZFS 6988 spa_sync() spends half its time in dmu_objset_do_userquota_updates

Using a benchmark which creates 2 million files in one TXG, I observe
that the thread running spa_sync() is on CPU almost the entire time we
are syncing, and therefore can be a performance bottleneck. About 50% of
the time in spa_sync() is in dmu_objset_do_userquota_updates().

The problem is that dmu_objset_do_userquota_updates() calls
zap_increment_int(DMU_USERUSED_OBJECT) once for every file that was
modified (or created). In this benchmark, all the files are owned by the
same user/group, so all 2 million calls to zap_increment_int() are
modifying the same entry in the zap. The same issue exists for the
DMU_GROUPUSED_OBJECT.

We should keep an in-memory map from user to space delta while we are
syncing, and when we finish, iterate over the in-memory map and modify
the ZAP once per entry. This reduces the number of calls to
zap_increment_int() from "number of objects modified" to "number of
owners/groups of modified files".

This reduced the time spent in spa_sync() in the file create benchmark
by ~33%, from 11 seconds to 7 seconds.

Upstream bugs: DLPX-44799
Ported by: Ned Bass <bass6@llnl.gov>

OpenZFS-issue: https://www.illumos.org/issues/6988
ZFSonLinux-issue: https://github.com/zfsonlinux/zfs/issues/4642
OpenZFS-commit: unmerged

Porting notes:
- Added curly braces around declaration of userquota_cache_t cache to
  quiet compiler warning;
- Handled the userobj accounting the same way it proposed in this path.

Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>

module/zfs/dnode.c

@@ -2001,9 +2001,6 @@ dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
 	boolean_t hole;
 	int i, inc, error, span;
 
-	dprintf("probing object %llu offset %llx level %d of %u\n",
-	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
-
 	hole = ((flags & DNODE_FIND_HOLE) != 0);
 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
 	ASSERT(txg == 0 || !hole);